Nvidia vs. Nouveau

echasslau · 2010-01-20 15:44:45

Hi folks.

I did some kind of benchmark to check the current progress of Nouveau development and I was nearly frustrated about the result.
The first part shows GtkPerf with 1000 cycles. The second one glxgears.

                         nvidia  nouveau

GtkEntry                   0,15     0,17
GtkComboBox                3,14     3,65
GtkComboBoxEntry           2,04     3,13
GtkSpinButton              0,34     0,55
GtkProgressBar             0,13     0,19
GtkToggleButton            0,61     1,20
GtkCheckButton             0,28     0,28
GtkRadioButton             0,36     0,53
GtkTextView - Add text    15,43    14,83
GtkTextView - Scroll       1,58     3,08
GtkDrawingArea - Lines     2,47    21,69
GtkDrawingArea - Circles  14,17    21,21
GtkDrawingArea - Text      4,23    14,61
GtkDrawingArea - Pixbufs   0,62     7,48
Total time                45,56    92,60

glxgears 1280x800           487       39
glxgears                   2600      611

Ok, I know the glxgears thing is not interessting at the moment because of the beta (or even alpha?) state of the DRM part. But even the 2D benchmark is not really competitive.
Is there any way to speed up the nouveau driver?

Last edited by echasslau (2010-01-20 15:47:36)

IncredibleLaser · 2010-01-20 17:10:36

glxgears is not interesting because it's not a benchmark.

flamelab · 2010-01-20 17:13:10

Nouveau will be speedy after the Gallium3D driver will be done. Now, it's just too slow, its only interesting feature is KMS, and ... nothing else.

ChemBro · 2010-01-20 17:15:39

IncredibleLaser wrote:

glxgears is not interesting because it's not a benchmark.

That's right, it's only one test. A benchmark consists of many test, so glxgears is part of his benchmark.

iBertus · 2010-01-20 17:26:04

The difference in speed while rendering 2D is still amazing. I didn't expect the binary blob to be so much faster while rendering the simple stuff.

How will gallium3d affect this performance?

tavianator · 2010-01-20 18:10:51

Gallium3d will have an X state-tracker, which basically replaces exa acceleration I believe.

Anyway, nouveau is slow because it's actually a challenging undertaking to write a device driver with no device documentation. It's not clear that it will ever be as fast as the blob, but improvements are constantly being made.

echasslau wrote:

Is there any way to speed up the nouveau driver?

Yep, by submitting patches

shining · 2010-01-26 23:30:01

echasslau wrote:

Hi folks.
I did some kind of benchmark to check the current progress of Nouveau development and I was nearly frustrated about the result.
The first part shows GtkPerf with 1000 cycles. The second one glxgears.
                         nvidia  nouveau

GtkEntry                   0,15     0,17
GtkComboBox                3,14     3,65
GtkComboBoxEntry           2,04     3,13
GtkSpinButton              0,34     0,55
GtkProgressBar             0,13     0,19
GtkToggleButton            0,61     1,20
GtkCheckButton             0,28     0,28
GtkRadioButton             0,36     0,53
GtkTextView - Add text    15,43    14,83
GtkTextView - Scroll       1,58     3,08
GtkDrawingArea - Lines     2,47    21,69
GtkDrawingArea - Circles  14,17    21,21
GtkDrawingArea - Text      4,23    14,61
GtkDrawingArea - Pixbufs   0,62     7,48
Total time                45,56    92,60

glxgears 1280x800           487       39
glxgears                   2600      611
Ok, I know the glxgears thing is not interessting at the moment because of the beta (or even alpha?) state of the DRM part. But even the 2D benchmark is not really competitive.
Is there any way to speed up the nouveau driver?

Can you try downclocking your card when running nvidia, to get the same gpu/mem clock speeds as with nouveau, and do a fair comparison ?
I would find it very interesting
You can use nvclock -i when using nouveau to see the speed. And you can also use nvclock to control the speeds when the nvidia driver loaded, to set the same default speed that nouveau gets.

If it explains the difference, the way to speed up will be to find out how to control the clock speeds.

echasslau · 2010-02-17 11:40:37

Here a real benchmark:
http://www.phoronix.com/scan.php?page=a … irst&num=1

shining · 2010-02-17 16:41:46

echasslau wrote:

Here a real benchmark:
http://www.phoronix.com/scan.php?page=a … irst&num=1

Well, he doesn't mention the clock speeds either, that's a shame ...
It probably would not change much but it could have explained some of the strange results he noticed, like the difference between the two cards using nouveau.

atomkarinca · 2010-02-17 20:59:16

I've been using the Nouveau drivers for over three months now and I can say I'm very happy with the performance. I don't use compositing and I don't play 3D games (only XMoto). I cannot notice any 2D performance differences compared to the blob but -as mentioned- 3D performance is very poor at the moment. Nouveau can run compositing to a certain extent -and surprisingly well- but I guess the real improvement will be with the widespread usage of the driver when Gallium3D comes into play.

Last edited by atomkarinca (2010-02-17 20:59:55)

echasslau · 2010-02-18 05:26:38

shining wrote:

echasslau wrote:
Here a real benchmark:
http://www.phoronix.com/scan.php?page=a … irst&num=1
Well, he doesn't mention the clock speeds either, that's a shame ...
It probably would not change much but it could have explained some of the strange results he noticed, like the difference between the two cards using nouveau.

The clock speed thing is on my ToDo list but not at the top...
(but my ToDo list is too long for my spare time i think)

Last edited by echasslau (2010-02-18 05:28:26)

shining · 2010-02-18 08:08:00

And it's indeed a more real benchmark but it's a 3d one.
When developing a driver (and any code), it doesn't make any sense to start with optimizing. You first start by implementing features while trying to keep stability. nouveau is still incomplete and unstable so that first stage is not done yet. Performance will come after.

A real 2d benchmark would be interesting though, I would be curious to see cairo perf trace results :
http://cworth.org/intel/performance_measurement/

echasslau · 2010-06-21 09:09:48

Hi folks.

I gave it a new try but again without checking the clocking speed.

                              2010-01-20        2010-06-21
                         nvidia  nouveau   nvidia  nouveau

GtkEntry                   0,15     0,17     0,19     0,21
GtkComboBox                3,14     3,65     4,15     3,98
GtkComboBoxEntry           2,04     3,13     2,88     3,62
GtkSpinButton              0,34     0,55     0,49     0,64
GtkProgressBar             0,13     0,19     0,35     0,82
GtkToggleButton            0,61     1,20     0,59     0,61
GtkCheckButton             0,28     0,28     0,39     0,40
GtkRadioButton             0,36     0,53     0,50     0,55
GtkTextView - Add text    15,43    14,83    15,85    15,85
GtkTextView - Scroll       1,58     3,08     2,48     2,55
GtkDrawingArea - Lines     2,47    21,69     5,46    16,98
GtkDrawingArea - Circles  14,17    21,21     7,05     8,31
GtkDrawingArea - Text      4,23    14,61     5,00    14,45
GtkDrawingArea - Pixbufs   0,62     7,48     0,66     7,26
Total time                45,56    92,60    46,06    76,23

glxgears 1280x800           487       39      497      207
glxgears                   2600      611     2750      931

The 3D part of nouveau is much faster, but I don't know why the 2D part is slower with both drivers. But it shows that nouveau is getting nearer to the performance of nvidia.

flamelab · 2010-06-21 10:40:09

2D is quite good actually. But in 3D, it is (in general) at least 20 times slower, for now.

echasslau · 2010-06-21 12:11:23

I've thought about it and I think the slower 2D part perhaps is because of the kernel.
My first test were with bfs kernel, now with the normal one.

Don't know why you say it's 20 times slower. For me it looks like:
497 / 207 = 2,4 and 2750 / 931 = 2,95

Last edited by echasslau (2010-06-21 12:15:10)

some-guy94 · 2010-06-21 17:46:56

glxgears is NOT a benchmark.
Not to mention that nouveau doesn't yet support nearly as many extensions as nvidia and is at opengl 2 while nvidia is at 4 (with compatible hardware)

echasslau · 2010-06-21 19:30:31

wikipedia:
In computing, a benchmark is the act of running a computer program, a set of programs, or other operations, in order to assess the relative performance of an object, normally by running a number of standard tests and trials against it.

what's wrong with my "set of programs"?

BTW: I did not mention the word benchmark!!!

Last edited by echasslau (2010-06-21 19:31:36)

Gusar · 2010-06-21 20:57:00

glxgears is an OpenGL program that reports FPS (frames per second) numbers. However, it is a very limited 'test'. Unlike most modern 3D games glxgears:
* has an extremely low vertex/polygon count
* does no texturing at all
* only simple, flat shading is used (except inside the hole in each gear is simple smooth shading)
* all vertex data is stored in a display list, so almost nothing passes between host CPU and video card once rendering is started. This mostly implies video card fill rate is limited. But, see next point.
* the default window size is 300x300, a large part of which is not even rendered into, so it's not even a good fill rate test
* the entire render step consists of only 21 OpenGL functions calls, of which only 6 are unique. This is not a very good OpenGL API stress test. Something like glean would be better.
So to summarize, glxgears only tests a small part of what you typically see in a 3D game. You could have glxgears FPS performance increase, but your 3D game performance decrease. Likewise, you could have glxgears performance decrease and your 3D game performance increase.

Source (first result on google for "glxgears is not a benchmark"): http://wiki.cchtml.com/index.php/Glxgea … _Benchmark

echasslau · 2010-06-21 20:59:30

who cares? it enough for a quick test...

some-guy94 · 2010-06-21 21:05:27

echasslau wrote:

who cares? it enough for a quick test...

A test, not a comparison.

Gusar · 2010-06-21 21:19:35

echasslau wrote:

who cares? it enough for a quick test...

Yeah, that tells you whether something is seriously wrong (when the number is like 20), or that you have vsync on (when the number is 60 or 75). It tells you absolutely nothing else.

onzlaught · 2010-07-02 03:37:24

echasslau wrote:

who cares? it enough for a quick test...

Absolutly right! Nouveau is slower... period! Why don't you just open a thread about glxgears? It's amazing how every forum is full of people how repeat the same thing they just learned over and over, i thinks it's because they know nothing else, they are unable to learn further and they want to show everybody what they know... nothing! Post a real benchmark if you want to contribute... the rest is already well known!

Cdh · 2010-07-02 04:44:42

Geforce FX 5200 with nouveau since 173xxx does not support 2.6.34. Flash Fullscreen Videos do not work. At least it does not crash but it only shows stange colors. Even the controls are not visible.
I tried neverball and it also only shows a strange colored empty window when clicking on start or options. It doesn't even work slow.

So much for the nvidia fanboys who always say that amd is slow to support new kernel and xorg. At least they release specs so the open source driver for cards that are not supported is good.

Gusar · 2010-07-02 07:02:17

Cdh wrote:

So much for the nvidia fanboys who always say that amd is slow to support new kernel and xorg.

Nice flame dude. A Nvidia user talking of Nvidia fanboys. It's kinda funny.

Last edited by Gusar (2010-07-02 07:05:12)

shining · 2010-07-02 07:22:48

Can you all stop being kids ? Otherwise that thread is going to close very soon.

No, glxgears is not a benchmark, so yes, if some people keep using it as such, then other people unfortunately need to keep repeating the information. Of course, just saying it's not a benchmark is not enough, you need to explain why, or better, just give a link that already explains well why :
http://wiki.cchtml.com/index.php/Glxgea … _Benchmark

And yes nvidia provides quality proprietary drivers and might be better than amd/ati in that regard. But they don't provide any specs or code, so that has some drawbacks too, especially with older cards that get quickly marked as "legacy" and no longer receive the attention they deserve. Nothing is ever all black&white.

Arch Linux

#1 2010-01-20 15:44:45

Nvidia vs. Nouveau

#2 2010-01-20 17:10:36

Re: Nvidia vs. Nouveau

#3 2010-01-20 17:13:10

Re: Nvidia vs. Nouveau

#4 2010-01-20 17:15:39

Re: Nvidia vs. Nouveau

#5 2010-01-20 17:26:04

Re: Nvidia vs. Nouveau

#6 2010-01-20 18:10:51

Re: Nvidia vs. Nouveau

#7 2010-01-26 23:30:01

Re: Nvidia vs. Nouveau

#8 2010-02-17 11:40:37

Re: Nvidia vs. Nouveau

#9 2010-02-17 16:41:46

Re: Nvidia vs. Nouveau

#10 2010-02-17 20:59:16

Re: Nvidia vs. Nouveau

#11 2010-02-18 05:26:38

Re: Nvidia vs. Nouveau

#12 2010-02-18 08:08:00

Re: Nvidia vs. Nouveau

#13 2010-06-21 09:09:48

Re: Nvidia vs. Nouveau

#14 2010-06-21 10:40:09

Re: Nvidia vs. Nouveau

#15 2010-06-21 12:11:23

Re: Nvidia vs. Nouveau

#16 2010-06-21 17:46:56

Re: Nvidia vs. Nouveau

#17 2010-06-21 19:30:31

Re: Nvidia vs. Nouveau

#18 2010-06-21 20:57:00

Re: Nvidia vs. Nouveau

#19 2010-06-21 20:59:30

Re: Nvidia vs. Nouveau

#20 2010-06-21 21:05:27

Re: Nvidia vs. Nouveau

#21 2010-06-21 21:19:35

Re: Nvidia vs. Nouveau

#22 2010-07-02 03:37:24

Re: Nvidia vs. Nouveau

#23 2010-07-02 04:44:42

Re: Nvidia vs. Nouveau

#24 2010-07-02 07:02:17

Re: Nvidia vs. Nouveau

#25 2010-07-02 07:22:48

Re: Nvidia vs. Nouveau

Board footer